Introduction

For our project, we decided to research the effectiveness of Government COVID-19 policies throughout the pandemic. There were several motivations for wanting to look into this topic. First and foremost, this topic was presented to us by Manhattan College for the annual Business Analytics Competition that they host. For this competition we were tasked with looking into different data sets provided to us by Bloomberg and Oxford, and drawing conclusions from these datasets. The guiding questions that we were presented with are as follows:

Not only were these topic interesting to us… we were also enticed by the chance at winning a cash prize if we placed top 3 in the final round of the competition. While we fell a little short of this goal, we still qualified for the final round, and placed 9th out of 34 teams competing.

Another motivation to look into this topic was how COVID-19 has greatly effected my personal life, and my partners personal life, as we are both students and athletes here at SUNY Geneseo. COVID-19 and the policies presented to us by the government have been a road bump in our busy college careers for about 2 years now, so looking into the best ways to approach a pandemic such as COVID-19 was immediately interesting to both of us. With that being said, part of the motivation to research this topic was to hopefully uncover indicators of successful practices regarding containing the spread of virus, minimizing the fatality of the virus, all while minimizing social and economic impacts. While there is no indication that a pandemic such as COVID-19 will hit again in the near future, it is best to at least begin to look into how we can contain such a catasrophe. There was no indication that COVID-19 would be so harmful, indicating that something like this could happen any day with no prior indication.

Overview of Modeling Techniques

Clutsering Analysis: Grouping data with similar outcomes. The goal of our clustering analysis was to discover countries that had similar government respomses to the COVID-19 Pandemic

Principal Compononent Analysis: The princiapl components describe the hyperellipsoid in N-space that roughly bounds the data.

Pooled Ordeinary Least Squrae: Normal liner regression model fitted using the OLS technique on a flattened version of the panel data set.

Fixed Effect Linear Regression: Statistical regression model which is useful for working with panel data. Panel data allows us to control for variables that we cannot observe or measre such a cultural factors or differences in social norms across countries

Data

The data we utilized for our research included two large datasets, one from Oxford and one from Bloomberg. The data presented to us from Oxford was titles “Oxford COVID Government Response Variables”. The dataset given to us from Bloomberg was a resiliency ranking that Bloomberg compiled of the top 53 ecoonomies in the world. We also utilized the Google COVID-19 Community Mobility Report to aid in our regression and analysis process.

Oxford

Oxford’s goal is to track COVID-19 policy data and compare each countrie’s responses/policies. The data includes 180 countries, utilizing 23 indicators. Collection of the data began January 1st of 2020.

Inidicators

There are a total of 21 live indicators in the dataset that are updated daily

Containment and Closure Policies: C1-C8

Economic Policies: E1-E4

Health System Policies: H1-H8

Vaccine Policies:: V1-V4

Policy Indices

Overall Government Response Index: Calculated using ordinal indicators

Containment and Health Index Combines lockdown restricstions and closure measures with health and variables such as testing policy, contact tracing, etc. Calculated using all ordinal containment and closure policy indicators and health system policy indicators.

Stringency Index: Measures the strictness of lockdown style. Calculated using all odrinal containment and closure policy indicators, plus an indicator recording public information campaigns

Economic Support Index: Measures income support and debt relief. Calculated using all ordinal economic policy indicators.

Risk of Openness Index: Based on the reccomendations set out by the World Health organization of measures that should be put in place before COVID-19 response policies can be safely relaxed

Calculation of Policy Indices

Policy indices are averages of the individual component indicators

Bloomberg’s Data

Bloomberg created resiliency rankings for the 53 largest economies in the world. A country’s rank is based on their success of controlling the virus with the least amount of social economic disruptions.

Reopening Group: Vaccine Doses per 100 people, Lockdown Severity, Flight Capacity, Vaccinated Travel Routes.

Covid Status Group: 1-Month Cases per 100k, 3-Month Case Fatality Rate, and Deaths Per 1 Million.

Quality of Life Group: Community Mobility, 2022 GDP Growth Forecast, Universal Healthcare Coverage, Human Development Index.

Ranking Method

Bloomberg used the “Max-Min” method where a score of 100 indicates the best performance, while 0 indicates the worst. The final score was then determined by averaging each country’s performances across all 12 indicators with equal weight.

Conceptual Framework

Variation in Government Response

The graphs shown above are representative of government stringency. The Gray lines represent each individual country’s government stringencies, and the blue is the trend line for the entire continent. As you can see, Africa, the Americas, and Asia all had similar negative trends regarding stringency. Europe did not follow this trend, as stringency fluctuated greatly over time. In Oceania, government stringency has steadily increased since late 2021 going into 2022.

We grouped the data by continent to emphasize the differences in reproduction rate around the world as the pandemic went on. We can see, for example, when comparing the Americas to Asia in January of 2022 the reproduction rate in Asia was much higher than that of America. This can be due to many different factors, but it may be due to level of government trust or differences in policy.

Above we see the Economic support overtime, again with the gray representing individual country’s results, and the blue line repreenting the trend for the entire continent. While there were no significant findings, it is worth noting that Africa had the least amount of economic support by a substantial amount.

Models

Clustering Using Principal Component Analysis

We utilized clustering analysis to visualize Bloomberg’s data. We did this in order to discover governments that had similar responses. We used Principal Component Analysis to characterize twelve Bloomberg indices into the two PCs that best explain the variance of the bloomberg indices.

Reproduction Rate Linear Regression

ReproductionRateit=β0+β1∗GovernmentPoliciesit+υi+λt+ϵit

Where:

υt=Country Fixed Effect

λt=Daily Fixed Effect

ϵt=Error Term

Death per Capita Linear Regression

DeathPerCapitait=β0+β1∗GovernmentPoliciesit+υi+λt+ϵit

Where:

υt=Country Fixed Effect

λt=Daily Fixed Effect

ϵt=Error Term

Description of Government Policy Variables

Keep in mind that the variables with the “lag2” prefix are lagged 1 week for the reproduction rate linear regression and 1 month for the death per capita regression.

We chose to use 1 week lagged variables to account for a possible delay in the ever changing reproduction rate and account for the incubation period of the virus. We chose 1 month lagged variables for the time it takes for policies to be implemented as well as the time it takes for the death per capita rate to change.

Please not that all variables with the Flag suffix are used to account for differences in policy characteristics. The flag is similar to a dummy variable with different characteristics depending on the particular policy.

lag2_StringencyIndex: Measures the strictness of lockdown style. It is calculated using all ordinal containment and closure policy indicators, plus an indicator recording public information campaigns.

lag2_E1_Income support: Recorded if the government is providing direct cash payments to people who lose their jobs or cannot work. This only includes payments to firms if explicitly linked to payroll or salaries. It is calculated on an ordinal scale. 0- No income support. 1- The government is replacing less than 50% of lost salary (or if a flat sum, it is less than 50% median salary). 2- The government is replacing 50% or more of lost salary (or if a flat sum, it is greater than 50% median salary).

Lag2_E1_Flag: Binary flag for sectoral scope. 0- formal sector workers only or informal sector workers only. 1- all workers.

lag2_E2_Debt/contract relief: Record if the government is freezing financial obligations for households (eg stopping loan repayments, preventing services like water from stopping, or banning evictions). It is calculated on an ordinal scale. 0- no debt/contract relief. 1- narrow relief, specific to one kind of contract. 2- broad debt/contract relief

lag2_H2_Testing policy: Records policies about testing for current infection (PCR tests) not testing for immunity (antibody test). It is calculated on an ordinal scale. 0- no testing policy. 1- only those who both (a) have symptoms AND (b) meet specific criteria (eg key workers, admitted to hospital, came into contact with a known case, returned from overseas). 2- testing of anyone showing Covid-19 symptoms. 3- open public testing (eg “drive through” testing available to asymptomatic people).

lag2_H3_Contact tracing: Records government policy on contact tracing after a positive diagnosis. It is calculated on an ordinal scale. 0- no contact tracing. 1- limited contact tracing; not done for all cases. 2- comprehensive contact tracing; done for all identified cases.

lag2_H4_Emergency investment in healthcare: Announced short term spending on healthcare system, eg hospitals, masks, etc. Note: only record amount additional to previously announced spending. Record monetary value in USD. 0- no new spending that day.

lag2_H5_Investment in vaccines: Announced public spending on Covid-19 vaccine development. Note: only record amount additional to previously announced spending. Record monetary value in USD. 0- no new spending that day.

lag2_H6_Facial Coverings: Record policies on the use of facial coverings outside the home. It is calculated on an ordinal scale. 0- No policy. 1- Recommended. 2- Required in some specified shared/public spaces outside the home with other people present, or some situations when social distancing not possible. 3- Required in all shared/public spaces outside the home with other people present or all situations when social distancing not possible. 4- Required outside the home at all times regardless of location or presence of other people.

lag2_H6_Flag: Binary flag for geographic scope. 0- Targeted. 1- General.

lag2_H7_Vaccination policy: Record policies for vaccine delivery for different groups. It is measured on an ordinal scale. 0- No availability. 1- Availability for ONE of following: key workers/ clinically vulnerable groups (non elderly) / elderly groups. 2- Availability for TWO of following: key workers/ clinically vulnerable groups (non elderly) / elderly groups. 3- Availability for ALL of following: key workers/ clinically vulnerable groups (non elderly) / elderly groups. 4- Availability for all three plus partial additional availability (select broad groups/ages). 5- Universal availability

lag2_H7_Flag: Binary flag for cost. 0- At cost to individual (or funded by NGO, insurance, or partially government funded). 1- No or minimal cost to individual (government funded or subsidized).

lag2_H7_Vaccination policy:lag2_Hy_Flag MobilityIndex: The Community Google Mobility Reports show movement trends by region, across different categories of places.

Results

Clustering Analysis Results

Figure A

This is the result of clustering on the Bloomberg Data. The X axis is representative of PC1, and the Y axis is representative of PC2

The value of PC1 is negatively correlated with vaccine doses per 100, and healthcare coverage. PC1 is positively correlated with 3-month case fatality rate.

PC2 is negatively correlated with flight capacity, and vaccinates travel roots, while it is positively correlated with 2021 GDP Growth Forecast.

Cluster 5 shows countries who had low vaccine doses per 100 and healthcare coverage, which is shown by PC1, and had moderate GDP growth forecast, which is shown by PC2

Linear Regression Reults

Reproduction Rate Results

Figure B

Figure C

Death Per Capita Results

Figure D

Figure E

Discussion of Findings

Reproduction Rate

In order to model our data in the best way possible, we ran five different linear regression models using reproduction rate as out independent variable. Our explanatory variables are 1 week lagged values of various policies implemented throughout the pandemic using economic and health policy variables. The first column is a pooled OLS model, followed by a fixed effect model on day, then a fixed effect on country, then both country and date fixed effect, and finally both fixed effects and the use of Google’s mobility index. By looking at the R-squared values of the regressions we ran on the reproduction rate, as we utilize fixed effects on date and country we were able to create a more explanatory model. It is also evident that including the mobility index makes our beta estimate more statistically significant. In the model with both daily and country fixed effets, along with with the google mobility index we are able to see that 37.9% of the variation in the reproduction rate can be explained by the varation with our independent variables. Figure C is representative of our beta estimates, which are graphed at the 95% condifence interval. The results of the beta estimates were a bit surprising. The stringency index was statistically significant from 0 and negative. This indicates that the more stringent the country’s policies, the more successful the country was at limiting the virus’ reproduction rate. With that being said, the magnitude of the negative was rather small. Another variable that was statistically significant and negative was vaccination policy. The magnitude of this negative correlation was much stronger. This indicates that as more people received the vaccine, the easier it was to control the spread of the virus. Within figure C there are variables that are “flagged” and not flagged. The flagged variables account for differences in similar policies amongst different countries, whereas the non flagged variables do not take minor differences into account. Keeping this in mind, when combining flagged and non flagged facial covering variables, the results were slightly negative, yet were not statistically significantly different from 0. Interestingly, investment in vaccines and emergency investment in healthcare had no effect on reproduction rate. Contact tracing was negative, but not significantly different from 0, as was testing policy. Debt/conract releif was siginificantly different from 0 and positive, meaning that for some reason these policies tended to increase the spread of COVID-19. This could be due to people having more money to spend, which would lead to more people going out in public to spend that money, thus contributing to the spread. Similarly, the income support variables combined to have a positive correlation, and was significantly different from 0.

Death per Capita

While looking at death per capita, we utilized the same 5 regression models (pooled ordinary least swaures, daily fized effect, country fixed effect, daily and country fixed effect, and both fixed effects with the google mobility index) as shown in figure D. As stated above, the addition of the fixed effects and google mobility increased our R-squared values. Again, our fifth model utilizng both fized effects and google moibility had the best R-squared value of 0.538m which menas that 53.8% of the variation in the death per capita can be explained by the variation in government policies. Figure E is the graph of our beta estimates. The stringency index was statstically significant and differnt from zero, and also positive. THe magnitude of this estimate is extremely small. The Vaccination policy variables (flagged and without flag) combined did not have a statistically significant beta estimate and was not different from 0. Facial coverings (again flagged and without flag) combined to have a negative beta estimate that was not statistically significant. Figure E also shows that investment in vaccines and emergency investment in healthcare did not have any impact on the death per capita rate. While the magnitude was still somewhat small, testing policy had the largest negative effect on death per capita, and was statistically significant, meaning that testing policies contributed to preventing COVID related deaths. Lastly, Debt/contract relief was a slight negative, but not statistically significant, and the flagged/without flag income support variables combined to create a beta estimate that is not statistically significant.

Conclusion

Keeping in mind the main purpose of researching government policies regarding COVID-19, which was to see what worked well when trying to contain the virus, there were two policies that stood above the rest in our models. Testing policies seemed to help the best to limit the death per capita of a country, whereas vaccination policy proved to be the best method of slowing down the reproduction rate of COVID-19.

A large limitation that both my partner and I encountered over the course of this research was the very large datasets that we had to work with. When it came time to both run the regressions and create visualizations, our computers processors were not strong enough. This made the project take a lot longer than we initially thought, and did not allow us to play around with the data as much as we would have liked to.

In regards to research in the future, there are a couple things, in my opinion, that could have made the data more accurate, thus making our models stronger. Looking at the Bloomberg data, we felt that the final resiliency rankings could have been approached using a weighted average of a country’s scores on the 11 indices, rather than a simple average of all of the scores. For example, a variable such as “universal healthcare coverage” does not hold as much importance, in terms of our research, as something such as “deaths per million”. Throughout our research, we were also aware that many countries failed to report on certain data, and obviously a lot of the numbers regarding COVID-19 deaths and cases were misrepresented throughout the world. There is no way to control this, because every country reported cases and deaths differently throughout the pandemic. The only true way to get the most accurate data would be to centralize the process of reporting cases/deaths (which is not reasonably possible) in the event that we encounter another pandemic.

Works Cited

Hale, Thomas, Jessica Anania, Noam Angrist, Thomas Boby, Emily Cameron-Blake, Martina Di Folco, Lucy Ellen, Rafael Goldszmidt, Laura Hallas, Beatriz Kira, Maria Luciano, Saptarshi Majumdar, Radhika Nagesh, Anna Petherick, Toby Phillips, Helen Tatlow, Samuel Webster, Andrew Wood, Yuxi Zhang, “Variation in Government Responses to COVID-19” Version 12.0. Blavatnik School of Government Working Paper. 11 June 2021. Available: https://www.bsg.ox.ac.uk/covidtracker

Thomas Hale, Noam Angrist, Rafael Goldszmidt, Beatriz Kira, Anna Petherick, Toby Phillips, Samuel Webster, Emily Cameron-Blake, Laura Hallas, Saptarshi Majumdar, and Helen Tatlow. (2021). “A global panel database of pandemic policies (Oxford COVID-19 Government Response Tracker).” Nature Human Behaviour. https://doi.org/10.1038/s41562-021-01079-8

World Economic Forum, “Global Risks Report 2022,” 11 Jan. 2022, https://www.weforum.org/reports/global-risks-report-2022/in-full/chapter-6-refreshing-resilience-from-covid-19-lessons-to-a-whole-of-society-response.

Zumel, Nina, et al. Practical Data Science with R. Manning Publications Co., 2020.